Rent a GPU, or pay per token?

The only question that matters is utilization. This calculator takes today's live GPU floors and API floors, your expected throughput, and tells you where the lines cross. Every field is editable. Our defaults are planning estimates, not benchmarks. Prefer a ready-made answer? The self-hosting cost table runs this math for every notable open model at once. Rather hand the whole decision off? Describe your workload and we'll send back a ranked API-vs-rent-vs-own plan, free while in beta.

Method: GPU count = VRAM needed ÷ card VRAM (rounded up, tensor parallelism assumed free; it isn't quite). Self-host $/1M = (GPU count × $/hr) ÷ (throughput × 3600 × utilization) × 10⁶. API comparison uses today's cheapest listed output price. Input-token costs, egress, storage and engineering time are not included.